Copilot commented on code in PR #12751: URL: https://github.com/apache/gluten/pull/12751#discussion_r3756580978
########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + + private def publishGlobalMinBfKey(entry: JoinInputEntry): Unit = { + val key = (entry.leafTableNames, entry.outputSignature) + val current = globalMinBfKeys.get(key) + if ((current eq null) || entry.bfExprKeys.size < current.size) { + globalMinBfKeys.put(key, entry.bfExprKeys) + } + } Review Comment: publishGlobalMinBfKey uses a non-atomic get+put on ConcurrentHashMap; concurrent rule invocations can overwrite a smaller baseline with a larger one, defeating the "historically smallest" invariant. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + + private def publishGlobalMinBfKey(entry: JoinInputEntry): Unit = { + val key = (entry.leafTableNames, entry.outputSignature) + val current = globalMinBfKeys.get(key) + if ((current eq null) || entry.bfExprKeys.size < current.size) { + globalMinBfKeys.put(key, entry.bfExprKeys) + } + } + + private def getGlobalBaselineBfKeys( + leafTables: Set[String], + outputSig: Seq[(String, DataType)], + entryBfKeys: Set[String]): Option[Set[String]] = { + val key = (leafTables, outputSig) + Option(globalMinBfKeys.get(key)) + .filter(g => g.subsetOf(entryBfKeys) && g != entryBfKeys) + } + + private def isPhysicalFilter(p: SparkPlan): Boolean = p match { + case _: FilterExec => true + case _: FilterExecTransformer => true + case _ => false + } + + private def filterCondition(p: SparkPlan): Expression = p match { + case FilterExec(cond, _) => cond + case FilterExecTransformer(cond, _) => cond + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def filterChild(p: SparkPlan): SparkPlan = p match { + case FilterExec(_, c) => c + case FilterExecTransformer(_, c) => c + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def copyFilterWith( + p: SparkPlan, + newCond: Expression, + newChild: SparkPlan): SparkPlan = p match { + case f: FilterExec => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case f: FilterExecTransformer => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def probeSignature(expr: Expression): (String, String, Option[Long]) = { + // 1. extract rawProbe (children(1)) + val rawProbeOpt: Option[Expression] = expr match { + case bf: BloomFilterMightContain => Some(bf.children(1)) + case veloxBf: VeloxBloomFilterMightContain => Some(veloxBf.children(1)) + case _ => None + } + + // 2. resolve probe attribute and optional hash seed + val (attrOpt, seedOpt) = rawProbeOpt match { + case Some(XxHash64(children, seed)) => + // collectFirst depth-first left-to-right picks up the first Attribute (incl. children.head) + (children.headOption.flatMap(_.collectFirst { case a: Attribute => a }), Some(seed)) + case Some(other) => + (other.collectFirst { case a: Attribute => a }, None) + case None => + (None, None) + } + + // 3. build the return tuple + attrOpt match { + case Some(a) => (a.name, a.dataType.simpleString, seedOpt) + case None => ("UNKNOWN", "UNKNOWN", seedOpt) + } + } + + private def exprKey(expr: Expression): String = { + val (name, dateType, seedOpt) = probeSignature(expr) + s"probe=$name:$dateType|seed=${seedOpt.getOrElse("NONE")}" + } + + private def unfoldQueryStages(plan: SparkPlan): SparkPlan = + plan.transformUp { + case s: ShuffleQueryStageExec => unfoldQueryStages(s.plan) + case b: BroadcastQueryStageExec => unfoldQueryStages(b.plan) + case r: ReusedExchangeExec => unfoldQueryStages(r.child) + } + + private def bloomFilterExprsWithHost( + plan: SparkPlan): Seq[(Expression, SparkPlan)] = { + val unfolded = unfoldQueryStages(plan) + unfolded.collect { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val conjuncts = splitConjunctivePredicates(condition) + conjuncts + .filter(isBloomFilter) + .map(bf => (bf, f)) + }.flatten + } + + private def stripExtraBloomFilters( + root: SparkPlan, + baselineBfKeys: Set[String]): (SparkPlan, Int) = { + var strippedCount = 0 + val rewritten = root.transformDown { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val child = filterChild(f) + val conjuncts = splitConjunctivePredicates(condition) + val remaining = conjuncts.filter { + expr => + val shouldStrip = isBloomFilter(expr) && !baselineBfKeys.contains(exprKey(expr)) + if (shouldStrip) { + strippedCount += 1 + } + !shouldStrip + } + remaining.reduceOption[Expression](And) match { + case Some(cond) if cond.fastEquals(condition) => f + case Some(cond) => copyFilterWith(f, cond, child) + case None => child + } + } + (rewritten, strippedCount) + } + + /** + * Extract the table name a leaf (Scan-like) SparkPlan reads from. This is used as one dimension + * of the exchange-reuse grouping key. Values are taken directly from strongly-typed fields + * (tableIdentifier.table / last segment of Table.name) -- no reliance on plan.simpleString and + * its truncation. Unsupported leaf types throw immediately. + */ + private def extractTableName(leaf: SparkPlan): String = { + def stripCatalog(name: String): String = { + val i = name.lastIndexOf('.') + if (i >= 0) name.substring(i + 1) else name + } + + leaf match { + case f: FileSourceScanExecTransformer => + f.tableIdentifier.map(_.table).get + case b: BatchScanExecTransformer => Review Comment: extractTableName calls Option.get on tableIdentifier; FileSourceScanExecTransformer.tableIdentifier can be None for file-based scans, which would throw and abort planning. This issue also appears on line 240 of the same file. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + + private def publishGlobalMinBfKey(entry: JoinInputEntry): Unit = { + val key = (entry.leafTableNames, entry.outputSignature) + val current = globalMinBfKeys.get(key) + if ((current eq null) || entry.bfExprKeys.size < current.size) { + globalMinBfKeys.put(key, entry.bfExprKeys) + } + } + + private def getGlobalBaselineBfKeys( + leafTables: Set[String], + outputSig: Seq[(String, DataType)], + entryBfKeys: Set[String]): Option[Set[String]] = { + val key = (leafTables, outputSig) + Option(globalMinBfKeys.get(key)) + .filter(g => g.subsetOf(entryBfKeys) && g != entryBfKeys) + } + + private def isPhysicalFilter(p: SparkPlan): Boolean = p match { + case _: FilterExec => true + case _: FilterExecTransformer => true + case _ => false + } + + private def filterCondition(p: SparkPlan): Expression = p match { + case FilterExec(cond, _) => cond + case FilterExecTransformer(cond, _) => cond + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def filterChild(p: SparkPlan): SparkPlan = p match { + case FilterExec(_, c) => c + case FilterExecTransformer(_, c) => c + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def copyFilterWith( + p: SparkPlan, + newCond: Expression, + newChild: SparkPlan): SparkPlan = p match { + case f: FilterExec => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case f: FilterExecTransformer => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def probeSignature(expr: Expression): (String, String, Option[Long]) = { + // 1. extract rawProbe (children(1)) + val rawProbeOpt: Option[Expression] = expr match { + case bf: BloomFilterMightContain => Some(bf.children(1)) + case veloxBf: VeloxBloomFilterMightContain => Some(veloxBf.children(1)) + case _ => None + } + + // 2. resolve probe attribute and optional hash seed + val (attrOpt, seedOpt) = rawProbeOpt match { + case Some(XxHash64(children, seed)) => + // collectFirst depth-first left-to-right picks up the first Attribute (incl. children.head) + (children.headOption.flatMap(_.collectFirst { case a: Attribute => a }), Some(seed)) + case Some(other) => + (other.collectFirst { case a: Attribute => a }, None) + case None => + (None, None) + } + + // 3. build the return tuple + attrOpt match { + case Some(a) => (a.name, a.dataType.simpleString, seedOpt) + case None => ("UNKNOWN", "UNKNOWN", seedOpt) + } + } + + private def exprKey(expr: Expression): String = { + val (name, dateType, seedOpt) = probeSignature(expr) + s"probe=$name:$dateType|seed=${seedOpt.getOrElse("NONE")}" + } + + private def unfoldQueryStages(plan: SparkPlan): SparkPlan = + plan.transformUp { + case s: ShuffleQueryStageExec => unfoldQueryStages(s.plan) + case b: BroadcastQueryStageExec => unfoldQueryStages(b.plan) + case r: ReusedExchangeExec => unfoldQueryStages(r.child) + } + + private def bloomFilterExprsWithHost( + plan: SparkPlan): Seq[(Expression, SparkPlan)] = { + val unfolded = unfoldQueryStages(plan) + unfolded.collect { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val conjuncts = splitConjunctivePredicates(condition) + conjuncts + .filter(isBloomFilter) + .map(bf => (bf, f)) + }.flatten + } + + private def stripExtraBloomFilters( + root: SparkPlan, + baselineBfKeys: Set[String]): (SparkPlan, Int) = { + var strippedCount = 0 + val rewritten = root.transformDown { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val child = filterChild(f) + val conjuncts = splitConjunctivePredicates(condition) + val remaining = conjuncts.filter { + expr => + val shouldStrip = isBloomFilter(expr) && !baselineBfKeys.contains(exprKey(expr)) + if (shouldStrip) { + strippedCount += 1 + } + !shouldStrip + } + remaining.reduceOption[Expression](And) match { + case Some(cond) if cond.fastEquals(condition) => f + case Some(cond) => copyFilterWith(f, cond, child) + case None => child + } + } + (rewritten, strippedCount) + } + + /** + * Extract the table name a leaf (Scan-like) SparkPlan reads from. This is used as one dimension + * of the exchange-reuse grouping key. Values are taken directly from strongly-typed fields + * (tableIdentifier.table / last segment of Table.name) -- no reliance on plan.simpleString and + * its truncation. Unsupported leaf types throw immediately. + */ + private def extractTableName(leaf: SparkPlan): String = { + def stripCatalog(name: String): String = { + val i = name.lastIndexOf('.') + if (i >= 0) name.substring(i + 1) else name + } + + leaf match { + case f: FileSourceScanExecTransformer => + f.tableIdentifier.map(_.table).get + case b: BatchScanExecTransformer => + stripCatalog(b.table.name) + case i: IcebergScanTransformer => + stripCatalog(i.table.name) + case f: FileSourceScanExec => + f.tableIdentifier.map(_.table).get + case b: BatchScanExec => + stripCatalog(b.table.name) + + case _ => + throw new IllegalStateException( + s"Unsupported leaf SparkPlan type for table extraction: " + + s"${leaf.getClass.getName} (nodeName=${leaf.nodeName})") Review Comment: extractTableName throws for unsupported leaf SparkPlan types. Because this rule is injected globally for Velox query-stage prep, an unexpected leaf type would fail query planning entirely; prefer a safe fallback so the rule becomes a no-op rather than aborting the query. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + + private def publishGlobalMinBfKey(entry: JoinInputEntry): Unit = { + val key = (entry.leafTableNames, entry.outputSignature) + val current = globalMinBfKeys.get(key) + if ((current eq null) || entry.bfExprKeys.size < current.size) { + globalMinBfKeys.put(key, entry.bfExprKeys) + } + } + + private def getGlobalBaselineBfKeys( + leafTables: Set[String], + outputSig: Seq[(String, DataType)], + entryBfKeys: Set[String]): Option[Set[String]] = { + val key = (leafTables, outputSig) + Option(globalMinBfKeys.get(key)) + .filter(g => g.subsetOf(entryBfKeys) && g != entryBfKeys) + } + + private def isPhysicalFilter(p: SparkPlan): Boolean = p match { + case _: FilterExec => true + case _: FilterExecTransformer => true + case _ => false + } + + private def filterCondition(p: SparkPlan): Expression = p match { + case FilterExec(cond, _) => cond + case FilterExecTransformer(cond, _) => cond + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def filterChild(p: SparkPlan): SparkPlan = p match { + case FilterExec(_, c) => c + case FilterExecTransformer(_, c) => c + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def copyFilterWith( + p: SparkPlan, + newCond: Expression, + newChild: SparkPlan): SparkPlan = p match { + case f: FilterExec => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case f: FilterExecTransformer => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def probeSignature(expr: Expression): (String, String, Option[Long]) = { + // 1. extract rawProbe (children(1)) + val rawProbeOpt: Option[Expression] = expr match { + case bf: BloomFilterMightContain => Some(bf.children(1)) + case veloxBf: VeloxBloomFilterMightContain => Some(veloxBf.children(1)) + case _ => None + } + + // 2. resolve probe attribute and optional hash seed + val (attrOpt, seedOpt) = rawProbeOpt match { + case Some(XxHash64(children, seed)) => + // collectFirst depth-first left-to-right picks up the first Attribute (incl. children.head) + (children.headOption.flatMap(_.collectFirst { case a: Attribute => a }), Some(seed)) + case Some(other) => + (other.collectFirst { case a: Attribute => a }, None) + case None => + (None, None) + } + + // 3. build the return tuple + attrOpt match { + case Some(a) => (a.name, a.dataType.simpleString, seedOpt) + case None => ("UNKNOWN", "UNKNOWN", seedOpt) + } + } + + private def exprKey(expr: Expression): String = { + val (name, dateType, seedOpt) = probeSignature(expr) + s"probe=$name:$dateType|seed=${seedOpt.getOrElse("NONE")}" + } + + private def unfoldQueryStages(plan: SparkPlan): SparkPlan = + plan.transformUp { + case s: ShuffleQueryStageExec => unfoldQueryStages(s.plan) + case b: BroadcastQueryStageExec => unfoldQueryStages(b.plan) + case r: ReusedExchangeExec => unfoldQueryStages(r.child) + } + + private def bloomFilterExprsWithHost( + plan: SparkPlan): Seq[(Expression, SparkPlan)] = { + val unfolded = unfoldQueryStages(plan) + unfolded.collect { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val conjuncts = splitConjunctivePredicates(condition) + conjuncts + .filter(isBloomFilter) + .map(bf => (bf, f)) + }.flatten + } + + private def stripExtraBloomFilters( + root: SparkPlan, + baselineBfKeys: Set[String]): (SparkPlan, Int) = { + var strippedCount = 0 + val rewritten = root.transformDown { + case f if isPhysicalFilter(f) => + val condition = filterCondition(f) + val child = filterChild(f) + val conjuncts = splitConjunctivePredicates(condition) + val remaining = conjuncts.filter { + expr => + val shouldStrip = isBloomFilter(expr) && !baselineBfKeys.contains(exprKey(expr)) + if (shouldStrip) { + strippedCount += 1 + } + !shouldStrip + } + remaining.reduceOption[Expression](And) match { + case Some(cond) if cond.fastEquals(condition) => f + case Some(cond) => copyFilterWith(f, cond, child) + case None => child + } + } + (rewritten, strippedCount) + } + + /** + * Extract the table name a leaf (Scan-like) SparkPlan reads from. This is used as one dimension + * of the exchange-reuse grouping key. Values are taken directly from strongly-typed fields + * (tableIdentifier.table / last segment of Table.name) -- no reliance on plan.simpleString and + * its truncation. Unsupported leaf types throw immediately. + */ + private def extractTableName(leaf: SparkPlan): String = { + def stripCatalog(name: String): String = { + val i = name.lastIndexOf('.') + if (i >= 0) name.substring(i + 1) else name + } + + leaf match { + case f: FileSourceScanExecTransformer => + f.tableIdentifier.map(_.table).get + case b: BatchScanExecTransformer => + stripCatalog(b.table.name) + case i: IcebergScanTransformer => + stripCatalog(i.table.name) + case f: FileSourceScanExec => + f.tableIdentifier.map(_.table).get + case b: BatchScanExec => + stripCatalog(b.table.name) + + case _ => + throw new IllegalStateException( + s"Unsupported leaf SparkPlan type for table extraction: " + + s"${leaf.getClass.getName} (nodeName=${leaf.nodeName})") + } + } + + /** + * Intermediate result shared between STEP1 and STEP4: everything that characterises one side of a + * join (one join input). + * + * @param joinInput + * the actual join-child subtree (used as IdentityHashMap key) + * @param hasBf + * whether any BloomFilter conjuncts exist under this subtree + * @param outputSignature + * output (name, type) pairs -- required equality dimension for exchange reuse + * @param leafTableNames + * set of all leaf table names reachable from this subtree -- required equality dimension for + * exchange reuse + * @param bfExprKeys + * unique BloomFilter identity keys encoded via exprKey() -- what STEP4 compares across sides + */ + private case class JoinInputEntry( + joinInput: SparkPlan, + hasBf: Boolean, + outputSignature: Seq[(String, DataType)], + leafTableNames: Set[String], + bfExprKeys: Set[String]) + + /** + * STEP4 output: one join-input side that needs BloomFilter stripping plus the baseline it will be + * stripped against. + * + * @param entry + * the side that has MORE BFs (the one carrying extras) + * @param baseline + * the reference side (its bfExprKeys are a strict proper subset of entry.bfExprKeys and + * strictly smaller) -- on stripping, only BFs whose exprKey is in baseline.bfExprKeys are kept, + * all the rest are removed. + */ + private case class StripTarget(entry: JoinInputEntry, baseline: JoinInputEntry) + + private def mkEntry(input: SparkPlan): JoinInputEntry = { + val withHost = bloomFilterExprsWithHost(input) + val bfExprKeys = withHost.map { case (bf, _) => exprKey(bf) }.toSet + val hasBf = bfExprKeys.nonEmpty + val sig = input.output.map(a => (a.name, a.dataType)) + val unfolded = unfoldQueryStages(input) + val leaves = unfolded.collectLeaves() + val leafNames = leaves.map(extractTableName).toSet + JoinInputEntry(input, hasBf, sig, leafNames, bfExprKeys) + } + + /** + * STEP1 helper: traverses ALL physical BinaryJoins inside the current physical plan (including + * any nested subqueries) and invokes mkEntry() separately on each join's left and right child. + * + * @return + * all produced JoinInputEntry instances. + */ + private def collectJoinInputs( + plan: SparkPlan): Seq[JoinInputEntry] = { + val allInputs = ArrayBuffer[JoinInputEntry]() + + def processJoinInput(input: SparkPlan): Unit = { + val entry = mkEntry(input) + allInputs += entry + } + + plan.collectWithSubqueries { + case b: BinaryExecNode => + processJoinInput(b.left) + processJoinInput(b.right) + } + + allInputs.toSeq + } + + private def isBloomFilter(expr: Expression): Boolean = expr match { + case _: BloomFilterMightContain => true + case _: VeloxBloomFilterMightContain => true + case _ => false + } + + override def apply(plan: SparkPlan): SparkPlan = { + + // ============================================================ + // STEP1 Collect JoinInputEntry: for every physical join + // (including nested subqueries), build one entry per + // join child: leaf-tables set, output-signature, and + // unique BloomFilter identity-key set. + // ============================================================ + val allJoinInputs = collectJoinInputs(plan) + val bfInputs = allJoinInputs.filter(_.hasBf) Review Comment: apply() always runs collectJoinInputs/mkEntry (which traverses join subtrees) before it knows whether BloomFilter predicates exist. For queries without runtime BloomFilters this adds avoidable planning overhead in every query stage; add a fast-path precheck that returns immediately when no BloomFilterMightContain/VeloxBloomFilterMightContain conjuncts are present. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + + private def publishGlobalMinBfKey(entry: JoinInputEntry): Unit = { + val key = (entry.leafTableNames, entry.outputSignature) + val current = globalMinBfKeys.get(key) + if ((current eq null) || entry.bfExprKeys.size < current.size) { + globalMinBfKeys.put(key, entry.bfExprKeys) + } + } + + private def getGlobalBaselineBfKeys( + leafTables: Set[String], + outputSig: Seq[(String, DataType)], + entryBfKeys: Set[String]): Option[Set[String]] = { + val key = (leafTables, outputSig) + Option(globalMinBfKeys.get(key)) + .filter(g => g.subsetOf(entryBfKeys) && g != entryBfKeys) + } + + private def isPhysicalFilter(p: SparkPlan): Boolean = p match { + case _: FilterExec => true + case _: FilterExecTransformer => true + case _ => false + } + + private def filterCondition(p: SparkPlan): Expression = p match { + case FilterExec(cond, _) => cond + case FilterExecTransformer(cond, _) => cond + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def filterChild(p: SparkPlan): SparkPlan = p match { + case FilterExec(_, c) => c + case FilterExecTransformer(_, c) => c + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def copyFilterWith( + p: SparkPlan, + newCond: Expression, + newChild: SparkPlan): SparkPlan = p match { + case f: FilterExec => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case f: FilterExecTransformer => + if ((newCond eq f.condition) && (newChild eq f.child)) f + else f.copy(condition = newCond, child = newChild) + case other => + throw new IllegalStateException(s"Not a physical filter: ${other.getClass.getName}") + } + + private def probeSignature(expr: Expression): (String, String, Option[Long]) = { + // 1. extract rawProbe (children(1)) + val rawProbeOpt: Option[Expression] = expr match { + case bf: BloomFilterMightContain => Some(bf.children(1)) + case veloxBf: VeloxBloomFilterMightContain => Some(veloxBf.children(1)) + case _ => None + } + + // 2. resolve probe attribute and optional hash seed + val (attrOpt, seedOpt) = rawProbeOpt match { + case Some(XxHash64(children, seed)) => + // collectFirst depth-first left-to-right picks up the first Attribute (incl. children.head) + (children.headOption.flatMap(_.collectFirst { case a: Attribute => a }), Some(seed)) + case Some(other) => + (other.collectFirst { case a: Attribute => a }, None) + case None => + (None, None) + } + + // 3. build the return tuple + attrOpt match { + case Some(a) => (a.name, a.dataType.simpleString, seedOpt) + case None => ("UNKNOWN", "UNKNOWN", seedOpt) + } + } + + private def exprKey(expr: Expression): String = { + val (name, dateType, seedOpt) = probeSignature(expr) + s"probe=$name:$dateType|seed=${seedOpt.getOrElse("NONE")}" + } Review Comment: Typo in exprKey(): local variable is named dateType but represents the probe data type; this makes the encoding harder to read and search for. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. Review Comment: This rule changes physical planning outcomes (strips BloomFilter predicates to recover ReusedExchange) but there is no targeted unit test validating the regression scenario. Consider adding a Velox AQE test that constructs a plan with asymmetric BloomFilterMightContain conjuncts across two otherwise-reuse-eligible join inputs and asserts a single ReusedExchange/scan after query-stage prep. ########## backends-velox/src/main/scala/org/apache/gluten/extension/RemoveBloomFilterToRecoverExchangeReuse.scala: ########## @@ -0,0 +1,472 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.gluten.extension + +import org.apache.gluten.execution.{BatchScanExecTransformer, FileSourceScanExecTransformer, FilterExecTransformer, IcebergScanTransformer} +import org.apache.gluten.expression.VeloxBloomFilterMightContain + +import org.apache.spark.internal.Logging +import org.apache.spark.sql.SparkSession +import org.apache.spark.sql.catalyst.expressions.{And, Attribute, BloomFilterMightContain, Expression, PredicateHelper, XxHash64} +import org.apache.spark.sql.catalyst.rules.Rule +import org.apache.spark.sql.execution.{BinaryExecNode, FileSourceScanExec, FilterExec, SparkPlan} +import org.apache.spark.sql.execution.adaptive.{BroadcastQueryStageExec, ShuffleQueryStageExec} +import org.apache.spark.sql.execution.datasources.v2.BatchScanExec +import org.apache.spark.sql.execution.exchange.ReusedExchangeExec +import org.apache.spark.sql.types.DataType + +import java.util.IdentityHashMap +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer + +/** + * Fixes a performance regression in TPC-DS Q24a/Q24b (and similar) queries on the Gluten Velox + * backend where asymmetric runtime BloomFilters injected by Spark cause the same large table (e.g. + * store_sales) to have different BF counts on the two join-input sides. This asymmetry makes + * canonicalized sameResult=false => ReusedExchange is disabled => the large table is scanned twice. + * + * The fix: on the join-input side that has MORE BloomFilters, precisely strip the extra BF + * conjuncts so that the canonicalized plans of the main query and the HAVING correlated + * scalar-subquery side become identical. Spark's native ReuseExchange rule then kicks in naturally, + * eliminating the duplicate scan. + * + * The apply() method runs in 5 phases: + * + * STEP1 Collect: traverse ALL physical joins (including those inside subqueries) in the current + * plan and build one JoinInputEntry per join child (leaf-tables-set, output-column signature, + * unique BF-keys set). STEP2 Publish: for every entry that carries BFs, publish its bfKeys into a + * cross-apply global pool. For each group (leafTables, outputSig) the pool retains the HISTORICALLY + * SMALLEST bfKeys set. STEP3 Group: cluster join inputs by (leafTables-set, output-signature) so + * that we only compare BF count asymmetry between join inputs that are actually eligible for + * exchange reuse. STEP4 Find asymmetry: within the same local group first look for a baseline whose + * bfKeys is a proper subset of entry.bfKeys and strictly smaller. If not found, fall back to the + * cross-apply global pool. If baseline exists and extraBf = entry.bfKeys -- baseline.bfKeys is + * non-empty, mark the entry for stripping. STEP5 Strip precisely: for each marked entry, walk + * top-down through all physical Filters under its subtree and drop ONLY those BF conjuncts whose + * exprKey is NOT in baseline.bfExprKeys. Leave isnotnull / other predicates intact. Finally graft + * the rewritten subtrees back into the original BinaryJoin's left/right children. + */ +case class RemoveBloomFilterToRecoverExchangeReuse(spark: SparkSession) + extends Rule[SparkPlan] + with PredicateHelper + with Logging { + + /** + * Cross-apply shared pool of the "historically smallest bfKeys set" per exchange-reuse group. + * + * Rationale: in Q24a the main query (bfCount=2) and the HAVING scalar subquery (bfCount=1) arrive + * in two completely separate apply() invocations because AQE splits them across different query + * stages. A purely local STEP4 would never see the smaller side as baseline -- the pool bridges + * that gap. + * + * Key = (leafTableNames, outputSignature): dimensions that uniquely define an exchange-reuse + * group. + * - leafTableNames: all leaf table names under this join input (e.g. {store_sales} or + * {store_sales,store_returns,store,item,customer} after multi-way joins) + * - outputSignature: sequence of (column-name, data-type) for the join input's output. Only + * join inputs sharing the exact same pair qualify for exchange reuse against each other. + * + * Value = Set[String] (bfExprKeys): the smallest bfExprKeys set historically published for this + * group. Encoding: see exprKey() -- "probe=<attr>:<type>|seed=<long|NONE>" On publish we only + * update if the new size is strictly smaller. On lookup only a strict proper-subset ("globalMin + * subsetOf entryBfKeys and globalMin != entryBfKeys") is returned as a valid baseline. + */ + private val globalMinBfKeys = + new ConcurrentHashMap[(Set[String], Seq[(String, DataType)]), Set[String]]() + Review Comment: globalMinBfKeys is a long-lived mutable map on a session-scoped rule instance; it can grow without bound in long-running SparkSessions (keys include potentially large leaf-table sets and output signatures). Consider scoping it to a single SQL execution (executionId) and/or bounding/evicting entries so it doesn't become an unbounded memory sink. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
